Bulk migration: manifest-driven runs, replayable failures, directory-level idempotency—no silent drops
`batch-vsd-docx` targets library- or folder-scale moves from .vsd to DOCX. What kills programs is partial success nobody notices, name collisions, and jobs you cannot rerun row-by-row. Freeze input manifest fields (relative path, repo version, case rules), log per-row status, duration, error code, and output hash, and keep read-only batch snapshots—no lockless multi-writer `out/`. Bucket oversized or OLE-heavy files separately so one monster file does not blow the whole SLA.
Bulk migration: golden and bad samples first, bucketed scale-out, daily manifest reconciliation
- Pilot the nastiest .vsd per business line plus known bad files; freeze server parameters and concurrency caps; publish the manifest schema and get data owners to sign—avoid everyone inventing their own spreadsheet mid-flight.
- Run with checkpoints every N rows: flush manifest and failure lists for resume; outputs must include repo name, path hash, or sequence—never flatten to bare filenames only.
- Reconcile manifest success counts against source inventories and object-store object counts; every delta needs a ticketed root cause (corrupt, encrypted, limit policy) before closing the batch—no “probably fine” closures.
VSD to DOCX (batch) FAQ: manifests, resume, overwrites, bad buckets
Morning after a night run, DOCX counts lag .vsd—which manifest columns do we read before blaming the engine?
Check each row’s `status`, skip reason, and whether an output path exists—often encoding, locks, or manual deletes, not random converter faults.
Same relative path across branches overwrites a shared bucket—how do we prove which commit produced the final DOCX?
Manifests must carry source checksum, git commit or repo version, and output hash; never treat unversioned `latest` folders as sole evidence.
A few huge .vsd files stall the batch—split buckets or lower global concurrency?
Create a large-file queue with higher timeouts; keep main concurrency healthy—global throttling stretches everyone’s wait without fixing single-file root causes.
Rerunning failures might overwrite hand-fixed DOCX—how do we prevent a second fight?
Mark manual_override rows in the manifest and change suffix or subdirectory; reruns skip unless a forced ticket says otherwise.
Compliance wants proof of a historical full-library snapshot—is the daily manifest enough?
No—also freeze object-store versions or read-only bucket snapshots plus a hash of the manifest itself; CSV alone can be edited after the fact.